Automated Capability Evaluation of Foundation Models

2025-05-22

AI Summary

Current foundation model evaluation relies on static, human-curated benchmarks, limiting comprehensive and dynamic characterization of model capabilities. To address this, we propose ACE, the first automated evaluation framework grounded in semantic space modeling and active learning. ACE first employs large language models to perform fine-grained semantic decomposition of domain-specific competencies; it then constructs a latent semantic space and leverages active learning to efficiently identify critical capability points; finally, it automatically generates diverse, targeted evaluation tasks. This paradigm eliminates dependence on fixed benchmarks and enables dynamic discovery of model weaknesses and failure modes. Experiments demonstrate that ACE substantially reduces human annotation effort while improving coverage breadth and diagnostic precision across capabilities. Crucially, it successfully uncovers key failure scenarios—such as reasoning inconsistencies and contextual misalignment—that are overlooked by conventional benchmarks, across multiple state-of-the-art foundation models.

Technology Category

• Natural Language Processing: Sentence-level Semantics, Textual Inference, etc.
• Machine Learning: Large Multimodal Models (LMMs)
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• User Modeling, Personalization and Recommendation: Metrics for user behavior and evaluating success

Abstract

Current evaluation frameworks for foundation models rely heavily on fixed, manually curated benchmarks, limiting their ability to capture the full breadth of model capabilities. This paper introduces Active learning for Capability Evaluation (ACE), a novel framework for scalable, automated, and fine-grained evaluation of foundation models. ACE leverages the knowledge embedded in powerful language models to decompose a domain into semantically meaningful capabilities and generate diverse evaluation tasks, significantly reducing human effort. To maximize coverage and efficiency, ACE models a subject model's performance as a capability function over a latent semantic space and uses active learning to prioritize the evaluation of the most informative capabilities. This adaptive evaluation strategy enables cost-effective discovery of strengths, weaknesses, and failure modes that static benchmarks may miss. Our results suggest that ACE provides a more complete and informative picture of model capabilities, which is essential for safe and well-informed deployment of foundation models.

Problem

• Overcoming limitations of fixed manual benchmarks for foundation models
• Automating scalable fine-grained evaluation using active learning
• Discovering model strengths and weaknesses missed by static methods

Innovation

• Active learning for automated model evaluation
• Decomposes domains into semantic capabilities
• Adaptive strategy prioritizes informative evaluations